Advanced machine learning techniques for State-of-Health estimation in lithium-ion batteries: A comparative study
Marek Sedlařík, Petr Vyroubal, Dominika Capková, Edin Omerdić, Mitchell Rae, Martin Mačák, Martin Šedina, Tomáš Kazda
Abstract
• Effective SOH estimation of Samsung INR18650-35E batteries using machine learning. • Comparative evaluation of SVR, FFNN, ANFIS, and GPR methods. • RMSE below 0.4 % achieved across all methods with half the training dataset. • FFNN achieved 0.39 % RMSE accuracy in B2 prediction based on B1 data. • Results emphasize the importance of method selection in battery monitoring. The accurate modeling and prediction of the State-of-Health (SOH) of lithium-ion (Li-ion) batteries are crucial for extending their lifespan, ensuring reliability, and minimizing the costs associated with extensive laboratory testing. This paper investigates the SOH estimation of Li-ion batteries utilizing advanced machine learning (ML) techniques. Specifically, 600 cycles were performed on Samsung INR18650-35E cells using the Constant Current Constant Voltage (CCCV) protocol. The input data for the ML methods were extracted from both charging and discharging cycles to achieve the best possible results. Data-driven models with different methodological foundations were used to predict SOH: Gaussian Process Regression (GPR), Support Vector Regression (SVR), and from the field of Artificial Neural Networks (ANN), Feed-Forward Neural Network (FFNN) and Adaptive Neuro-Fuzzy Inference System (ANFIS), which utilizes fuzzy logic. The input features for the ML methods were analyzed using Pearson Correlation Analysis (PCA), and additional inputs for the ANFIS method were selected using Exhaustive Search (ES) to identify the optimal combination of inputs with the lowest Root Mean Square Error (RMSE). The individual ML methods were evaluated on datasets of various sizes using the features with the highest correlation to SOH and the full set of features to detect overfitting. Further experiments explored the dependency of RMSE on the amount of training data, and SOH estimation of one battery was performed using training data from another. Overall, experiments show that nearly all methods achieved RMSE below 0.5% for SOH estimation, with SVR proving the most stable technique and ANFIS excelling with meticulously optimized configurations.